Content label establishment method based on electronic commerce data and social media marketing content data

Abstract

The present invention relates to the field of content marketing. Disclosed is a content label establishment method based on electronic commerce data and social media marketing content data. Aiming at the problems of inaccuracy and low efficiency in existing labeled entity extraction results, the present invention provides the content label establishment method based on electronic commerce data and social media marketing content data, comprising: constructing a commodity knowledge map; constructing a content category database; and carrying out information extraction on the content category database to construct a category content label tree, and on the basis of the category content label tree, labeling data in the content category database. According to the present invention, when information extraction is carried out on the content category database, a RaNER model is used to identify a concrete entity, then a large language model is used in combination with an information extraction type prompt and a chain of though summary type prompt to identify an abstract entity, and results extracted by a plurality of model entities are fused, such that the problems of inaccurate entity extraction and identification and incomplete entity recall of a single model are effectively solved, and high entity extraction accuracy and efficiency are achieved.

Description

A method for establishing content tags based on e-commerce data and social media marketing content data Technical Field

[0001] The present invention belongs to the field of content marketing technology, and more specifically, relates to a method for establishing content tags based on e-commerce data and social media marketing content data. Background Art

[0002] Data labeling is a crucial step in data processing. Labeling not only improves data readability and usability, but also facilitates data retrieval, analysis, and utilization. Data labeling is typically done manually. While highly accurate, it requires significant labor and time, especially when dealing with large data volumes. Another approach is automated annotation, which utilizes computer technology to label data. While highly efficient, it is prone to mislabeling or omissions. As is well known, entity extraction is required before labeling. This is highly complex, not only due to the complexity of documents but also due to the inherent difficulties of language understanding. In some scenarios, even humans struggle. For example, language contains implicit professional background knowledge and implicit or explicit context. The same text can express completely different concepts, while the same concept can be expressed in a variety of ways. These combined factors have made language understanding the crown jewel of artificial intelligence. Entity extraction from text inevitably requires language understanding, and the effectiveness of entity extraction depends on the level of language understanding. Entity extraction is directly linked to the accuracy of labeling.

[0003] Corresponding improvements have also been made to the above-mentioned problems, such as the Chinese patent application number CN111985237A, which was published on November 24, 2020. The patent discloses an entity extraction method, device and equipment, which includes: obtaining information to be processed, and the information to be processed includes at least two groups of interactive information data of interactions between a first object and a second object; performing multi-label classification and identification on each interactive information data in the information to be processed, so as to label the information to be processed with different functional types and obtain label labeling results; according to the label labeling results, performing entity extraction on the message to be processed in layers; the label labeling results at least include entity labels for indicating that the corresponding interactive information data contains entity blocks. The shortcoming of this patent is that although it can effectively identify complete entity information, the recognition accuracy is poor.

[0004] Another example is Chinese patent application number CN116610869A, published on August 18, 2023. This patent discloses a recommended content management method and device, electronic device, and storage medium. The method includes: sending a content recommendation request to a server, wherein the content recommendation request is used to request the server to generate and return recommended content for multiple display areas of a display interface; receiving multiple recommended contents returned by the server, and a data tag for each of the multiple recommended contents, wherein the data tag for each of the multiple recommended contents is obtained by performing NER entity extraction on each recommended content on the server, and the multiple recommended contents correspond one-to-one to the multiple display areas; deduplicating the multiple recommended contents based on the data tag of each recommended content in the multiple recommended contents, and exposing at least one recommended content remaining after the deduplication process to the corresponding display area. The shortcoming of this patent is that the data tags obtained only by NER entity extraction have certain deviations and poor accuracy.

[0005] Summary of the Invention

[0006] 1. Problems to be solved

[0007] To address the inaccurate and inefficient results of existing labeled entity extraction, the present invention provides a method for establishing content tags based on e-commerce data and social media marketing content data. This method uses the RaNER model to identify concrete entities when extracting information from a content category database. It then uses a large language model combined with information extraction prompts and thought chain summary prompts to identify abstract entities. The results of entity extraction from multiple models are then integrated, effectively compensating for the inaccurate extraction and incomplete entity recall caused by single-model entity extraction. This method achieves high entity extraction accuracy and efficiency.

[0008] 2. Technical solution

[0009] To solve the above problems, the present invention adopts the following technical solutions.

[0010] A method for establishing content tags based on e-commerce data and social media marketing content data includes the following steps:

[0011] S1: Build a product knowledge graph;

[0012] S2: Build a content category database; obtain a media content database, use the product knowledge graph to filter out the category-related content data of each product from the media content database, and build a content category database;

[0013] S3: Extract information from the content category database to construct a category content label tree, and label the data in the content category database according to the category content label tree. This step specifically includes the following steps:

[0014] S31: First, use the named entity recognition model to extract person entities, category entities, brand entities, and product attribute entities from the content category database. Then, use the large language model combined with information extraction prompts and thought chain summary prompts to perform semantic recognition on the content category database and extract person entities, network hot word entities, user pain point entities, product feature entities, and applicable entities. Finally, perform entity fusion to obtain the final entity.

[0015] S32: The final entity is converted into a word vector using a text vectorization model. A clustering algorithm is then used to obtain several categories of word vectors. The words in each category are then grouped into one or more tags using a large language model. The keyword types output by the large language model are used to construct a tree-structured content tag tree, along with keywords for each tag.

[0016] S33: Label the content text in the content category database according to the category content label tree.

[0017] Furthermore, the specific process of step S33 is as follows:

[0018] When labeling the content text, determine whether the content text has been subjected to entity extraction; if entity extraction has been performed, the entity words become candidate tags; if entity extraction has not been performed, perform entity extraction and add the tags corresponding to the identified entity words to the candidate tag set;

[0019] And for the keywords or regular expressions corresponding to each tag in the tag tree, keyword matching and regular expression matching are used, and the matched tags are also added to the candidate tag set of the content text;

[0020] Use the large language model to use the discriminant prompt to judge all candidate tags that have been screened out for the content text to determine whether the candidate tags match the meaning of the corresponding content text; if they match, confirm the candidate tag; if not, make corrections.

[0021] Furthermore, after the tree-structured content tag tree is constructed, manual review and correction of the content tag tree is also included.

[0022] Furthermore, the step S1 specifically includes the following steps:

[0023] Collect product information from various e-commerce platforms to form a product database;

[0024] The RaNER model is used to perform entity recognition on the commodity text information in the commodity database to construct a commodity knowledge graph; the structure of the commodity knowledge graph is in the form of a triple of (entity, relationship, entity).

[0025] Furthermore, the step S2 specifically includes the following steps:

[0026] S21: Collecting content information from various social media platforms to form a media content database;

[0027] S22: Use the product knowledge graph to perform text matching on the text information in the media content database to establish a preliminary screening database for product categories;

[0028] S23: Convert the image type content and video type content in the primary screening database into text content respectively;

[0029] S24: Perform fine screening and classification on the primary screening database to determine whether the text content is related to the data of the product category.

[0030] Furthermore, in step S21, the media content database only stores original text description information: for graphic content, its content title, content text and picture link content are stored; for video content, its content title and video link content are stored.

[0031] Furthermore, in step S23, for the image content, OCR technology is used to convert the text in the image into text content; for the video content, the video is first converted into an image, and the OCR technology is used to convert the text in the image into text content; and ASR technology is used to convert the video voice into text content, and the two text contents are combined to obtain the final text content of the video content.

[0032] Furthermore, the fine screening classification in step S24 specifically includes: after matching the brand and category keywords of the preliminary screening data in the preliminary screening database, if both the brand and category entities are matched, it is determined that the text content is related to the data of the product category; if only the brand or category keywords are matched, the large language model is used to classify the text content.

[0033] 3. Beneficial effects

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] (1) The present invention constructs a commodity knowledge graph and uses the entities and relationships in the commodity knowledge graph to construct a content category database from the media content database, so that the construction efficiency of the content category database is fast and work efficiency is effectively improved; after the construction of the content category database is completed, information is extracted from it to construct a content tag tree, and when performing information extraction, the RaNER model is used to identify concrete entities, and then the large language model is used in combination with the information extraction prompt and the thinking chain summary prompt to identify abstract entities. Different types of entities are identified and extracted using different models, which effectively compensates for the problems of inaccurate extraction and recognition and incomplete entity recall caused by entity extraction using a single model; finally, the content text is labeled using the content tag tree; the entire process is simple and not cumbersome, the entity extraction accuracy is high, and the overall labeling accuracy is high; at the same time, the efficiency is high, reducing labor costs and time costs;

[0036] (2) When labeling a content text, the present invention first determines whether entity extraction should be performed on the content text to improve work efficiency and save time; entity words that have been entity extracted become candidate tags, and entity words that have not been entity extracted become candidate tags after entity extraction. At the same time, the obtained candidate tag set is judged to avoid the possibility of potential errors, and the recall rate is improved as much as possible (reducing omissions). At the same time, the semantic understanding ability of the large language model is used to filter out entities that are matched by keywords but are semantically incorrect, and to filter out entity words that do not exist in the original text that may be given by the large language model as much as possible, thereby improving the precision rate;

[0037] (3) When constructing a content category database, the present invention includes first performing a preliminary screening of the media content database to obtain a preliminary screening database, then converting the pictures and video contents in the preliminary screening database into text content, and finally performing fine screening and classification to obtain a final content category database; the process of first performing a low-cost and rapid preliminary screening, narrowing the data range, and then performing fine screening reduces costs while improving efficiency while ensuring overall accuracy; and only original text description information is stored in the media content database, which greatly reduces storage costs;

[0038] (4) When converting video content into text content, the present invention first extracts frames from the video at a certain time interval, converts the video into a group of pictures, performs OCR processing on the pictures to convert them into text, and uses the position and size information of the text to filter out minor text information in the background as much as possible, while retaining important content text such as subtitles; then uses ASR to convert the video voice into text. The combination of OCR and ASR is conducive to a more complete analysis of the language information in the video. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] FIG1 is a schematic flow diagram of the present invention;

[0040] Figure 2 is a schematic structural diagram of a commodity knowledge graph;

[0041] FIG3 is a schematic diagram of the process of constructing a content tag tree;

[0042] FIG4 is a schematic diagram of a process of tagging content text. DETAILED DESCRIPTION

[0043] The present invention is further described below with reference to specific embodiments and accompanying drawings.

[0044] Please refer to Figures 1 to 3, Figure 1 is a flow chart of this application; Figure 2 is a schematic structure diagram of the product knowledge graph; Figure 3 is a flow chart of building a content tag tree; Figure 4 is a flow chart of tagging content text.

[0045] As shown in FIG1 , a method for establishing content tags based on e-commerce data and social media marketing content data includes the following steps:

[0046] S1: Build a product knowledge graph;

[0047] In one example, step S1 includes the following steps:

[0048] S11: Collect product information from various e-commerce platforms to form a product database. Product information includes platform, product ID, product name, specifications, product parameters, etc. Therefore, an example of product data in the product database is shown in Table 1 below:

[0049] Table 1: Product data table

[0050] S12: Use the RaNER model to perform entity recognition on the product text information in the product database to obtain brand, category, product attribute entities and related keywords, and use the co-occurrence relationship of the entities in the products to construct a product knowledge graph; the structure of the product knowledge graph is in the form of a triple (entity, relationship, entity);

[0051] For example, RaNER is a named entity recognition model for the Chinese e-commerce domain. The structure of the RaNER model can be found in the following paper: Improving Named Entity Recognition by External Context Retrieving and Cooperative Learning. The entity types that the RaNER model can recognize are shown in Table 2 below:

[0052] Table 2 Entity types that can be recognized by the RaNER model

[0053] For example, using the RaNER model to recognize the product title "Brand A Retinol Night Cream 2.0 Morning C Night A Firming Anti-Wrinkle" can produce the following results, as shown in Table 3:

[0054] Table 3 Recognition results

[0055] Among them, product_core product is a sub-category of commodities; based on the recognized entity words and the relationship between the co-occurrence of entity words in commodities, a commodity knowledge graph is constructed. An example of constructing a commodity knowledge graph based on the above entity recognition results is shown in Figure 2;

[0056] S2: Build a content category database; obtain a media content database, use the product knowledge graph to filter out the category-related content data of each product from the media content database, and build a content category database;

[0057] Specifically, step S2 includes the following steps:

[0058] S21: Collect and integrate content information from various social media platforms to form a media content database. The content on social media primarily includes text, images, and videos. In this step, the media content database stores only original text description information: for text, images, and images, the content title, content text, and image link content are stored; for video, the content title and video link content are stored. The media content database stores only original text description information and does not include multimedia data such as images and videos, thereby significantly reducing storage costs.

[0059] In one example, the content data of the media content database may be shown in Table 4 below:

[0060] Table 4 Content data table of media content database

[0061] S22: Use the product knowledge graph to perform text matching on the text information in the media content database to establish a preliminary screening database for product categories. It should be noted that since the content information on social media includes topics from various fields, both related to and unrelated to the products, it is necessary to screen out content related to products in a certain category. At the same time, since the media content database contains massive amounts of data (more than 1 billion items), and the content types include not only text but also video types, which need to be converted to text for further processing. The computing cost of video-to-text conversion is high and time-consuming, and using AI models to classify massive amounts of data is also high. Therefore, text matching is first performed on the text information to obtain a preliminary screening database. Low-cost and rapid preliminary screening narrows the data scope, ensuring both efficiency and reducing costs.

[0062] In the initial screening phase, keywords of various entities in the product knowledge graph, including keywords of related entities such as brands, categories, and attributes, are used to perform text matching on text information in the media content database. This step matches only the video titles of the video content, thereby quickly filtering out content data related to products in a certain category from massive content data and establishing an initial screening database for that category. A category initial screening database can be established for all target categories required by the business, such as a category initial screening database for beauty and personal care products.

[0063] S23: Convert the image-type content and video-type content in the preliminary screening database into text content respectively; since the preliminary screening database in step S22 contains image-type content and video-type content, only the text information such as the title is determined. In order to further analyze the text information related to the products contained in the images and videos, it is necessary to convert the content in the images and videos into text;

[0064] Specifically, for image content, OCR technology is used to convert the text in the image into text content; for video content, OCR technology and ASR technology are used to convert the video into text content: before OCR processing, the video is framed at a certain time interval, such as one frame per second, and the video is converted into a group of images. OCR technology is used to convert the text in the image into text content for this group of images, and the position and size information of the text is used to filter out minor text information in the background as much as possible, while retaining important content text such as subtitles; and ASR technology is used to convert the video voice into text content, thereby adding an (OCR text, ASR text) field to each video content; the two text contents are combined to obtain the final text content of the video content; since the text and voice on the screen in the video each contain some language information, and each also lacks some language information, and OCR and ASR technologies may have certain error rates, converting a video into text using OCR and ASR at the same time and using them together later is conducive to a more complete analysis of the language information in the video;

[0065] S24: Perform fine screening and classification on the primary screening database to determine whether the text content is related to the data of the product category;

[0066] After obtaining the text content of the image content and video content in the preliminary screening database in step S23, it is necessary to further determine whether the content is related to the product data of this category; because the preliminary screening uses keywords to quickly match to narrow the data range in massive data and recall potentially relevant content, there will be ambiguous keywords. For example, the milk brand "Guangming" may match "sunny" with keywords. Therefore, it is necessary to further refine the classification of the category preliminary screening database. The traditional method is to manually label the data first, and then train a supervised text classification model to determine whether the content text is related to a certain category of goods. Due to the large number of product categories that need to be processed, the cost of manually labeling data and then training the classification model is relatively high.

[0067] In step S24, the fine screening classification specifically includes: after matching the primary screening data in the primary screening database with brand and category keywords, if both brand and category entities are matched, it is determined that the text content is related to the product category data; if both brand and category entities are matched, no model judgment is required, thereby increasing the speed and reducing computing power consumption; if only brand or category keywords are matched, due to the high possibility of ambiguity, a large language model is used to construct a text classifier using prompt words such as "Does the above text describe a skin care product?" to classify the content text into the corresponding category, and then construct a content category database to complete the construction; Prompt is a prompt text that guides the model to generate a targeted answer;

[0068] S3: Extract information from the content category database to construct a category content label tree, and label the data in the content category database according to the category content label tree. This step specifically includes the following steps as shown in Figure 3:

[0069] S31: First, use the RaNER model to extract person entities, category entities, brand entities, and product attribute entities from the content category database. Then, use the large language model combined with information extraction prompts and thought chain summary prompts to perform semantic recognition on the content category database and extract person entities, network hot word entities, user pain point entities, product feature entities, applicable population entities, and applicable scenario entities. Finally, perform entity fusion to obtain the final entity.

[0070] In this step, we take the following example: For example, converting short video content into text:

[0071] OCR text:

[0072] Hello everyone, I'm **~ A beauty blogger with four years of experience. I highly recommend this makeup brand for new and year-end makeup. From their foundation cream to their entire product line now, I'm completely in love. Without further ado, let's apply it on bareface and see the results. Primer is a must. Then, I use a self-bandaging cushion. I use 574, which is suitable for healthy skin tones. You can choose 564 Natural, which has a three-layer high-elasticity powder puff. It applies makeup quickly and adheres perfectly, as if I were born with good skin. It even flatters my dry skin. Finally, I use a water-based loose powder to set my makeup. This water-based loose powder has a very fine texture, sets makeup long-lastingly, and replenishes moisture. It's the perfect base for a foundation cushion. The makeup effect and longevity are excellent when used together. Furthermore, Brand B's products are tested on allergy-sensitive skin and have no ingredients listed for caution. Hurry and get yours while you can!

[0073] ASR text:

[0074] Hello, everyone, I am **, a beauty blogger with 4 years of experience in the industry. I strongly recommend this makeup brand. You don't have to use it. It's the only choice for new year makeup. From the very beginning, I bought their foundation cream to the entire product line now. I'm really in love with it. Without saying too much, let's put on our bare eyes to see the effect directly. Primer is a must. This tender and non-greasy quality is very easy to spread, and it almost hides the pores. There is a 7% point that I use 574, which is suitable for fair skin. For healthy skin, you can choose 564. For natural color, the matching three-layer high-elastic powder puff is fast and adheres to the skin. It doesn't get caked and will make your skin look white. It looks like all the good reviews. It makes my dry skin adhere to the skin. Finally, I use water powder to set the makeup. It is the perfect makeup base CP for electricity. The combination of makeup effect and long-lasting makeup secret plus the whole foundation is very clear, adhered, clean and natural, and I am in good condition and want to buy it quickly;

[0075] In an example, using RaNER to extract entity effects:

[0076] Retouching_Other: Makeup

[0077] Modified_Other: Blogger

[0078] Scope of application_Applicable scenario: Year

[0079] Product_Core product: Foundation cream

[0080] Scope of application_Other: No makeup

[0081] Scope of application_Other: face

[0082] Product_Core product: Primer

[0083] Style_Other: Self-bandage

[0084] Products_Core products: air cushion

[0085] Dimensions: 574

[0086] Color_Color: Skin color

[0087] Color_Other: Healthy color

[0088] Color: 564 natural color

[0089] Dimensions_Other: Three layers

[0090] Modification_Product attributes: High elasticity

[0091] Product_Core product: powder puff

[0092] Scope of application_Other: dry skin

[0093] Product_Core product: The texture of water loose powder is particularly fine

[0094] Function and efficacy: Makeup setting

[0095] Function and efficacy: for skin

[0096] Function and effect: replenish moisture

[0097] Products_Core products: Loose powder

[0098] Products_Core products: air cushion

[0099] Brand: Brand B

[0100] Scope of application_Applicable people: partners

[0101] It should be noted that RaNER is relatively accurate in extracting category information, brand information, and some product attribute-related information. Its extraction performance on some semantically rich content texts is average, and its ability to understand the semantics of long content texts is insufficient. For example, the meaning of words such as "desert dry" and "for skin" is unclear; therefore, a large language model (LLM model) is used for entity recognition, and multiple prompts are used for extraction. The LLM model can be used for various NLP tasks, including entity recognition; the LLM model is an artificial intelligence model designed to understand and generate human language. They are trained on large amounts of text data and can perform a wide range of tasks, including text summarization, translation, sentiment analysis, etc.; however, there are still some problems when using LLM extraction, such as incomplete entity recall. This solution uses multiple types of prompts for entity recognition on a piece of content text, and fuses the results;

[0102] Among them, the information extraction prompt, such as "extract keywords such as product selling points, user pain points, etc. from the text", uses the LLM model combined with the information extraction prompt. The extraction results are shown in Table 5:

[0103] Table 5 LLM model combined with information extraction prompt extraction results

[0104] The extraction results of LLM demonstrate its semantic understanding ability. For example, “good skin born by mom”, “dry skin in the desert”, and “long-lasting makeup” are all valid content label phrases, and the corresponding types are also relatively accurate.

[0105] It should be noted that the LLM model has the problem of generating "hallucinations", such as "Internet hot words-Internet hot words". There are also problems such as some valid meanings not being extracted, or the extracted keywords being too long.

[0106] Therefore, we further use the chain of thought summary prompt to extract and calibrate entities through the LLM model. Chain of Thought is a prompt word pattern of the LLM large language model. By breaking down a multi-step reasoning problem into multiple intermediate steps, it guides the large model to gradually give the answer to the question, thereby improving the accuracy and explainability of the large model's answer to the question. This solution uses the chain of thought prompt to guide the large model to answer the question step by step and summarize it into the final required content label. This prompt method is called the chain of thought summary prompt. Unlike the information extraction prompt method that allows the LLM to give the answer directly, using the chain of thought to guide the LLM to gradually provide the intermediate process and summarize the answer can effectively avoid the problem of hallucination generation in a single answer, and can also perform more complex tasks. The prompt structure is such as "First find the text in the above text that describes the product selling points, user pain points..., and then summarize them into content labels..." The result example of using the LLM large model combined with the chain of thought summary prompt is:

[0107] {

[0108] "Product Selling Point - Popular Amway": "Popular Amway foundation brand, different choices for new year makeup",

[0109] "Product Selling Point - Full Line": "From the very beginning I bought their foundation cream to now their full line of products",

[0110] "Product selling point - essential primer": "Primer is a must-have texture."

[0111] Product Selling Point - Multiple Color Options: "I use the 574 self-bandage cushion, which is suitable for healthy skin tones. You can choose the 564 natural color."

[0112] Product Selling Point - Fitting Effect: "The three-layer high-elastic powder puff applies makeup quickly and fits perfectly. It's like my mother had perfect skin. It softens my dry skin perfectly."

[0113] "Product selling point - fine texture and hydration": "The texture of the water loose powder is extremely fine, long-lasting and sets makeup while also replenishing moisture to the skin."

[0114] "Product Selling Point - Excellent Lasting Power": "The perfect combination of foundation and air cushion for a flawless finish and excellent lasting power."

[0115] "Product selling point - tested for allergies and sensitive skin": "Brand B's products feature ingredients tested for allergies and sensitive skin, and beauty repair should be used with caution."

[0116] "Ingredients - Beauty Repair Ingredients": "Brand B's products feature ingredients that have been tested for allergies and sensitive skin and should be used with caution."

[0117] "User Pain Point - Moisturizing for Dry Skin": "It softens my dry skin perfectly."

[0118] "Applicable scene - before makeup use": "The texture of primer is a must-have",

[0119] "Applicable scenario - setting makeup": "Finally, use water and loose powder to set the makeup."

[0120] "Suitable skin type - dry skin": "It soothes my dry skin perfectly",

[0121] "Applicable Role - Makeup Tools": "The three-layer high-elastic powder puff applies makeup quickly and smoothly, just like naturally good skin. It even smooths out my dry skin."

[0122] "Sales Promotion - Sensitive Skin Tested, Beauty Repair Ingredients": "Brand B's products feature ingredients tested on sensitive skin and recommended for use with caution."

[0123] }

[0124] Therefore, the thought chain summary prompt enables LLM to summarize some complex statements. Compared with the information extraction prompt, its effect has some advantages and disadvantages, as shown in Table 6:

[0125] Table 6 Description of different prompt types

[0126] The combined use of these two methods can achieve better label word extraction results than a single prompt, further ensuring the accuracy of entity extraction and, therefore, the accuracy of subsequent content label tree construction. Furthermore, the entity words extracted from the content text using the RaNER model and the LLM model, including the correspondence between content IDs and entity word IDs, are stored in the word extraction database table. This result can be reused in the subsequent content labeling process according to the structured label tree.

[0127] S32: The final entity (entity word and entity word type, plus the attribute words extracted from the product, such as entity words of the function and efficacy type) is converted into a word vector through a text vectorization model; then, a clustering algorithm is used to obtain several categories of word vectors (word vectors with high similarity are clustered into one category); due to certain limitations of the clustering algorithm, after simple word vector clustering, words in the same category may still correspond to different meanings. Therefore, the semantic understanding ability of the large language model is then used to summarize the words in each category into one or more labels, and the keyword type output by the large language model is used to construct a tree-structured content tag tree and the keywords of each tag; the content tag tree is shown in Table 7 below as an example of a structured tag tree for beauty and personal care products:

[0128] Table 7 Structured label table of beauty and personal care products

[0129] Once the content tag tree is constructed, it undergoes manual review and correction. This includes adding tags, modifying tags, merging tags, and adding keywords or regular expressions to tags. After the tag tree is manually reviewed and corrected, the content text needs to be re-labeled according to the revised tag tree due to changes in the tag tree. This manual review and correction step further ensures the accuracy of tag creation.

[0130] S33: Label the content text in the content category database according to the category content tag tree; traditional methods for labeling content text according to the content tag tree include keyword matching, using a word segmentation model, and using a NER (named entity recognition) model. Keyword matching can accurately identify the keywords in the tag tree contained in the content text, and it is very fast, but there are some problems such as text ambiguity that cannot be distinguished, the problem of negative meanings being identified as positive meanings, and the problem that different expressions of the same meaning may not be recognized due to an incomplete vocabulary. Using a word segmentation model has a high dependence on the vocabulary, and some words that are not in the vocabulary may be misclassified, and the different word segmentation granularity may affect the subsequent tag recognition. Using a NER named entity recognition model, you can use a trained open source model or supervised training for a specific field, but the NER model may also have certain accuracy issues, or the cost of manually annotated data used to train the model to improve the accuracy in a specific field is high, or the recognized entity words cannot be mapped to the existing tag tree. Using the LLM large model, entity recognition can also be performed on content text. Compared with traditional methods, the large model has better semantic understanding capabilities and is more likely to correctly identify the semantics of a word in context. However, there is a certain probability that the required entity will be missed in each answer to the question. There is also the possibility of "hallucination" problem of the large model, where the entity given does not exist in the original text. Therefore, in this step, this application adopts an integrated method of rules, RaNER, LLM model and multiple prompts to give full play to the semantic understanding ability of the LLM model while avoiding some of the shortcomings of the LLM model.

[0131] As shown in FIG4 , the specific process of step S33 is as follows:

[0132] S331: When labeling the content text, determine whether entity extraction has been performed on the content text (i.e., whether extraction has been performed using the RaNER and LLM models). If entity extraction has been performed, the entity words become candidate tags. If entity extraction has not been performed, entity extraction is performed (including RaNER entity recognition and recognition using the LLM large model using the information extraction prompt and the thought chain summary prompt). The tags corresponding to the recognized entity words are added to the candidate tag set.

[0133] S332: keyword matching and regular expression matching are performed on the keywords or regular expressions corresponding to each tag in the tag tree, and the matched tags are also added to the candidate tag set of the content text;

[0134] S333: The candidate tag set obtained through the first two steps may contain potential errors. Tags identified by keywords may have some semantic errors, such as actually having a negative meaning, such as recognizing "comfortable" in "not very comfortable", or containing keywords but actually having another meaning, such as recognizing "dry skin" (a description of skin quality) in "wipe dry skin", or recognizing "olive skin" as a description of skin color or olive skin. For tags identified using the LLM model, semantic understanding errors are generally rare, but there is a certain probability that there are actually entities that are not mentioned in the original text (the "hallucination" problem of the large model). It is also possible that entity words, keywords, and other steps are recognized that are omitted in the LLM model output. Therefore, this step uses the large language model to use the discriminant prompt to judge all candidate tags that have been filtered out of the content text. The prompt word, such as "Judge whether the content text mentions the following content tags...", determines whether the candidate tag matches the meaning of the corresponding content text; if it matches, the candidate tag is confirmed; if not, it is corrected. In this way, while improving the recall rate (reducing omissions) through various means, the semantic understanding ability of the LLM model is used to filter out entities that are matched by keywords but are semantically incorrect as much as possible, and to filter out entity words that do not exist in the original text that the LLM model may give in the previous link as much as possible, thereby achieving the effect of simultaneously improving the precision and recall rates.

[0135] After identifying the content tags of each content text (title, video / image OCR text, video ASR text, etc.) of each content ID, the tag ID identified for the content ID is stored in the database table.

[0136] After determining whether all content in the content database is related to certain product categories, performing tag identification and storing the tag identification results, we can count the various interactive indicators of each content tag, including the number of likes, comments, and reposts, so as to evaluate the effectiveness of the content tags that appear in the marketing content of a certain product category and provide guidance on content tags for customers when producing product marketing content.

[0137] For the product database, the same structured tag tree and the same method are used for tag identification. In addition to using RaNER to identify entities in the products in the previous steps, some content tag words may not be recognized. For example, in the earlier product example "A Brand Bi-A Alcohol Night Cream 2.0 Retinol Cream Morning C Evening A Firming Anti-Wrinkle", RaNER did not recognize the content tag "Morning C Evening A". Combined with the constructed tag tree, the tag can be identified under the condition that the "Morning C Evening A" tag is already included in it. The tag identification results for all products in the product database are also stored in the product tag result database table.

[0138] The examples described in the present invention are merely descriptions of the preferred embodiments of the present invention and are not intended to limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various modifications and improvements made to the technical solutions of the present invention by engineers and technicians in this field should fall within the scope of protection of the present invention.

Claims

1. A method for establishing content tags based on e-commerce data and social media marketing content data, characterized by: The steps include: S1: Build a product knowledge graph; S2: Build a content category database; obtain a media content database, use the product knowledge graph to filter out the content data related to the category of each product from the media content database, and build a content category database; S3: extract information from the content category database to build a category content label tree, and label the data in the content category database according to the category content label tree; this step specifically includes the following steps: S31: First, use the named entity recognition model to extract person entities, category entities, brand entities, and product attribute entities from the content category database; then use the large language model combined with the information extraction prompt and the thinking chain summary prompt to perform semantic recognition on the content category database and extract person entities, network hot word entities, user pain point entities, product feature entities, and applicable entities; finally, perform entity result fusion to obtain the final entity; S32: Convert the entity words of the final entity into word vectors through the text vectorization model; then obtain several categories of word vectors through the clustering algorithm; then summarize the words in each category into one or more tags through the large language model, and use the keyword types output by the large language model to build a content tag tree with a tree structure and the keywords of each tag; S33: Tagging the content text in the content category database according to the category content tag tree.

2. The method for establishing content tags based on e-commerce data and social media marketing content data according to claim 1, characterized in that: The specific process of step S33 is as follows: When labeling the content text, determine whether the content text has been subjected to entity extraction; if entity extraction has been performed, the entity word becomes a candidate tag; if entity extraction has not been performed, after entity extraction, the tag corresponding to the identified entity word is added to the candidate tag set; And for the keywords or regular expressions corresponding to each tag in the tag tree, keyword matching and regular expression matching are used, and the matched tags are also added to the candidate tag set of the content text; Use the large language model to use the discriminant prompt to judge all the candidate tags that have been screened out for the content text to determine whether the candidate tags match the meaning of the corresponding content text; if they match, confirm the candidate tag, if not, modify it.

3. The method for establishing content tags based on e-commerce data and social media marketing content data according to claim 2, characterized in that: When the tree-structured content tag tree is constructed, manual review and correction of the content tag tree is also included.

4. The method for establishing content tags based on e-commerce data and social media marketing content data according to claim 1, characterized in that: The step S1 specifically includes the following steps: Collect product information from various e-commerce platforms to form a product database; The RaNER model is used to perform entity recognition on commodity text information in the commodity database to construct a commodity knowledge graph; the structure of the commodity knowledge graph is in the form of a triple of (entity, relationship, entity).

5. A method for establishing content tags based on e-commerce data and social media marketing content data according to claim 1 or 4, characterized in that: The step S2 specifically includes the following steps: S21: Collecting content information of various social media to form a media content database; S22: Use the product knowledge graph to perform text matching on the text information in the media content database to establish a preliminary screening database for product categories; S23: Convert the image type content and video type content in the preliminary screening database into text content respectively; S24: Perform fine screening and classification on the primary screening database to determine whether the text content is related to the data of the product category.

6. The method for establishing content tags based on e-commerce data and social media marketing content data according to claim 5, characterized in that: In step S21, the media content database only stores original text description information: for graphic content, the content title, content text and picture link content are stored; for video content, the content title and video link content are stored.

7. The method for establishing content tags based on e-commerce data and social media marketing content data according to claim 5, characterized in that: In step S23, for the image content, the OCR technology is used to convert the text in the image into text content; for the video content, the video is first converted into an image, and the OCR technology is used to convert the text in the image into text content; and the ASR technology is used to convert the video voice into text content, and the two text contents are combined to obtain the final text content of the video content.

8. The method for establishing content tags based on e-commerce data and social media marketing content data according to claim 5, characterized in that: The fine screening and classification in step S24 specifically includes: after matching the initial screening data in the initial screening database with brand and category keywords, if both brand and category entities are matched, it is determined that the text content is related to the data of the product category; if only brand or category keywords are matched, the large language model is used to classify the text content.

Images